InspectorSelectField's placeholder is unreachable at all 45 call sites

[os-justin]

Filed by the domain:ui PM seat (session_01YBWFb5YgMU5dw8p2VKj16S) on behalf of the objectui#6830 dev, who measured it while producing PR #8431 and could not file it (search_issues rate-limited). ⛔ Not claimed.

⚠️ This blocks the select half of objectui#6830's ruled direction A. That card's standing ruling is "show the declared default, do not write it"; showing it as placeholder text is impossible until this is fixed. Filed separately on purpose — a 61-call-site primitive change does not belong inside a flow-node card.

Measured

InspectorSelectField (packages/app-shell/src/views/metadata-admin/inspectors/_shared.tsx) declares placeholder = '—' and passes it to SelectValue placeholder={placeholder} on the trigger. It can never show. It bridges the caller's value through toInner(value ?? ''), so empty/undefined becomes the sentinel __inspector_select_none__ — a controlled value matching no SelectItem, which Radix renders as nothing. Radix shows a placeholder only for an undefined value.

Measured by rendering the primitive directly:

value undefined                          → trigger textContent ''        (not '—')
value '' + explicit placeholder 'PICKME' → trigger textContent ''        (not 'PICKME')
value matching an option                 → that option's label           ← the control: the field IS lit

Both negatives are therefore real absences, not a dead render.

61 call sites across the metadata-admin designer ⇒ every empty select renders blank instead of its hint.

Note on the existing test

_shared.select.test.tsx case 3 asserts an explicit placeholder is absent and attributes it to the '' option matching. The assertion is correct; its stated reason is not the operative one. Whoever fixes this should correct the reason rather than delete the case — the case becomes the ablation.

Fix shape (not ruled)

The mechanism is clear (a sentinel controlled value where Radix needs undefined), but the repair has a choice in it: pass undefined to Radix when there is no value and handle "no selection" some other way, or keep the sentinel and render the placeholder outside SelectValue. Both change what an empty select looks like across the whole designer, so the PR wants a before/after of a representative sample, not just the primitive's own test.

⚠️ Same shape as objectui#6830 itself — a declared prop that reaches no rendered output — one level down in a shared primitive.

Related

objectui#6830 / PR #8431 (where it was measured; its ruled direction A depends on this) · objectui#3912 (the same box, validation affordance)

Dedup

⚠️ Not run, declared rather than hidden. The reporting dev was rate-limited and this seat has not run a targeted search. No dedup claim is made. Suggested query for a triager: InspectorSelectField placeholder SelectValue sentinel radix never renders.

[os-justin]

Claim: session session_01YBWFb5YgMU5dw8p2VKj16S · branch claude/issue-8450-inspector-select-placeholder

PM dispatch (domain:ui seat). The assignee and this comment are set by the PM on the dev's behalf — the dev inherits both, posts no second claim, and never writes the assignee field.

Decision frame. The mechanism is measured and not in doubt: the field bridges the caller's value through toInner(value ?? ''), so empty/undefined becomes a sentinel matching no SelectItem, and Radix shows a placeholder only for an undefined value. What is not settled, and is yours to measure and propose: which repair — pass undefined to Radix and handle "no selection" another way, or keep the sentinel and render the placeholder outside SelectValue. Measured cost is the whole basis; no axis analysis expected. If the two are equivalent on cost, say so and pick, with the reason.

⚠️ 61 call sites. Whichever repair you choose changes what an empty select looks like across the whole designer. The PR needs a before/after of a representative sample — not just the primitive's own test — and the sample has to be chosen by measurement (which call sites pass an explicit placeholder, which rely on the '—' default, which are never empty in practice), not by convenience.

⚠️ The existing test is a trap and a gift. _shared.select.test.tsx case 3 asserts an explicit placeholder is absent and attributes it to the '' option matching. The assertion is correct; its stated reason is not the operative one. Correct the reason rather than delete the case — that case becomes your ablation.

⚠️ Ask whether an implementation strictly worse than the bug would pass: a field that renders the placeholder always, including when a value is selected, satisfies "the placeholder appears when empty". Pin the selected-value case as the non-regression half.

Context: this unblocks the select half of objectui#6830's ruled direction A (its boolean half landed as PR #8473). ⛔ Do not implement that half here — this card is the primitive only.

---

Generated by Claude Code

[os-justin]

os-dev-report

{
"issue": 8450,
"status": "done",
"branch": "claude/issue-8450-inspector-select-placeholder",
"pr": "#8494",
"premise_still_valid": true,
"summary": "InspectorSelectField now renders its placeholder itself, on the narrow state that means nothing is selected: no value AND no option standing for none. The mechanism in the card was confirmed exactly - the sentinel bridge guarantees Radix never sees the '' or undefined it needs for SelectValue's placeholder, so a controlled value matching no SelectItem painted a blank trigger. Two logic lines plus a JSX swap in packages/app-shell/src/views/metadata-admin/inspectors/_shared.tsx; data-placeholder is set on the trigger so Shadcn's data-[placeholder]:text-muted-foreground still greys the empty state, which is legitimate because SelectPrimitive.Trigger spreads caller props after its own. Branch pushed, draft PR #8494 open against main, 4 files, +269/-27. HEAD is 7d664d4, a merge of origin/main (PR #8473 landed mid-run and touched the same test file; conflict was one docblock paragraph, resolved by taking their boolean text verbatim and re-applying my select correction).",
"repair_chosen": "B - keep the sentinel, render the placeholder outside SelectValue. NOT equivalent on cost to A, contrary to the dispatch's 'if the two come out equivalent, say so and pick'. Both need the identical predicate and change the identical set of call sites, so they cost the same to write; A additionally makes the Radix Select UNCONTROLLED for as long as the field is empty, because undefined is how Radix spells uncontrolled and there is no other way to say 'controlled, nothing selected'. Measured against the real Select with a console.warn spy: (1) empty to first selection, undefined -> 'profile', logs 1 warning 'Select is changing from uncontrolled to controlled' - that is every call site's first interaction; (2) value untouched, options arriving late and gaining a none row, logs the same warning - the async pickers useMetaOptions / datasetOptions / chartXOptions / fieldOptions all populate after mount, so controlled-ness would flip on data arrival alone; (3) while uncontrolled, useControllableState stores Radix's own value via setUncontrolledProp, so a pick the owner declines to persist stays on screen (read from the installed source, not rendered). B measured 0 warnings on both transitions, selected label survives the empty-to-selected swap, and the ''-option round trip is intact. Two mechanics B needed: the placeholder cannot be SelectValue's children, because SelectItemText only portals the selected label into the value node when valueNodeHasChildren is false; and the placeholder is rendered INSTEAD of SelectValue, not beside it, so the trigger keeps two flex children under justify-between.",
"sample_choice_and_before_after": "The dispatch's sampling question has a degenerate answer: 0 of the 45 non-test call sites pass an explicit placeholder. The only explicit placeholder in the whole tree is 'Pick one' inside _shared.select.test.tsx. So the useful split is a different one, produced by a brace-balanced JSX scan that resolves each value expression through its const initializer and each options identifier to its array literal: bucket A, options offer a '' row so empty IS a selection, 10 sites, unchanged; bucket B, can be empty with no '' row, 13 sites, blank becomes em-dash; bucket C, value always resolves to a non-empty key, 22 sites, unchanged. (The static rule put 14 in B; ActionDefaultInspector.tsx:530 is a false positive, its || 'expression' makes it never empty, moved to C by hand. Four C members are empirically rather than provably non-empty: DatasetDefaultInspector d.type and m.aggregate, ObjectFieldInspector:645, ViewColumnInspector:270.) The sample takes one site per bucket and runs them through the REAL inspectors, four components, six triggers - before then after, trigger textContent: FlowNodeConfigField.tsx:225 http_request Method, empty to em-dash; AppNavInspector.tsx:391 nav item whose type infers null, empty to em-dash; PageBlockInspector.tsx:545 object-form Mode, empty to em-dash; PageBlockInspector.tsx:545 object-form Layout, empty to em-dash; ObjectFieldInspector.tsx:911 field Group with a '' row, '— No group —' both sides; FlowNodeInspector.tsx:401 node type, 'http_request' both sides. The last two are the sample's own controls and did not move.",
"shared_select_case_3": "Kept, assertions unchanged, stated reason corrected, and a third assertion added (the trigger text is the option label ALONE, so an implementation rendering label plus placeholder also fails). The old comment read the placeholder's absence as 'the '' option matched'. True of the first assertion; the second held for a different reason entirely - before this change the placeholder was absent in EVERY state, matching option or not, so the case could not have told a working field from a broken one. The comment now says exactly that and points at the neighbouring case where the same props with the '' row stripped render 'Pick one', which is what makes the absence mean what it always claimed.",
"always_placeholder_caricature": "MEASURED, not argued, and the measurement corrects what I would have reported. Mutated the trigger to render the placeholder unconditionally, keeping data-placeholder conditional so the caricature is tested in its most favourable form. Result on _shared.select.test.tsx: 5 failed | 5 passed (10). The case that actually refuses it, and the quietest one, is the PRE-EXISTING 'shows the selected non-empty option label on the trigger' - 'the selected option label reaches the DOM: expected null not to be null'. Also red: case 3, 'the "" option is a real selection and its label wins: expected null not to be null'; 'does NOT render the placeholder when a value is selected'; 'does NOT render the placeholder when "" is itself an offered option'; and 'flips both ways as the value comes and goes', which fails on 'expected em-dash to be Metadata'. Applying the objectui#8474 correction to my own pin: the most technical-looking case, 'carries Radix's own data-placeholder flag in the empty state only', SURVIVES the caricature - the attribute logic is independent of the text rendering, so that showpiece cannot distinguish the fix from its worst caricature at all. I had reasoned before measuring that it would refuse; it does not. The load-bearing refusals are the plain positive text-equality assertions on a selected value, not the attribute case.",
"ablation": "Ablated the READ SITE - restored the pre-fix trigger (a bare SelectValue placeholder element under an unconditional SelectTrigger) on top of the committed implementation, direction predicted before running. Hashes: HEAD blob ccdf0f887245e15fd17e16a12752a8582f65f069; on disk before mutation the same, so the tree was clean; on disk after mutation 91998e75aa76c3c53bf440a0f33bb5d4627832a0, differing. Both grep directions with the matched line printed: removed marker 'data-placeholder={showPlaceholder' 1 to 0, line 297 shown; injected marker 'SelectValue placeholder={placeholder}' 0 to 1, line 295 shown. No rebuild is owed - the pins reach the module by the relative source import from './_shared', not through a package exports entry, so no dist copy can serve a stale green. Result 7 failed | 16 passed (23). Red by name: the five _shared.select.test.tsx rows 'renders the declared em-dash default when the value is undefined', 'renders the declared em-dash default for an empty-string value', 'renders the CALLERs placeholder - the prop that reached no output before', 'flips both ways as the value comes and goes', 'carries Radix's own data-placeholder flag in the empty state only'; plus FlowNodeInspector.declaredDefault.test.tsx 'select: an unset key draws the placeholder, not the declared GET' and 'a server-published default is derived into the field and still not rendered'. Green throughout: both 'does NOT render' rows and all three 'empty-value option' rows - that asymmetry is the point. Restored BY STATE: git hash-object back to ccdf0f887245e15fd17e16a12752a8582f65f069, equal to the HEAD blob, and git diff HEAD empty; re-run after restore 23 passed (23). Script carried trap on EXIT INT TERM with absolute paths. The same discipline was used for the caricature run above, restored the same way. The _shared.tsx blob at the final merged HEAD 7d664d4 is still ccdf0f887245e15fd17e16a12752a8582f65f069, so the ablation describes the shipped head exactly.",
"objectui_3912": "Read in full. NO collision and NO pre-emption. It is about BlockPropField's text branch in previews/block-config.ts carrying placeholder as its only capability, and wants pattern / validate / description so identifier-valued curated fields (record:details sections[].name, page:tabs items[].key, page:accordion items[].value) get mechanical rejection instead of a hint string. That placeholder is InspectorTextField's, which renders fine as the native placeholder attribute on an input element - a different atom from the one this card repairs, and #3912's complaint is that placeholder is overloaded to carry validation semantics, not that it fails to render. This PR touches neither block-config.ts nor InspectorTextField and adds no capability to BlockPropField. The single point of contact is that kind:'select' block props at PageBlockInspector.tsx:545 now show em-dash when unset, which affects nothing #3912 proposes. It is labelled pm:on-hold and unassigned.",
"changeset": "patch for @object-ui/app-shell, in .changeset/8450-inspector-select-placeholder.md. Argument: InspectorSelectField is NOT exported from any package index - grep of the barrels shows zero re-exports, it is internal to the metadata-admin designer - so no published API moves and no consumer's types change. What changes is rendered output inside a shipped app surface, which is what a patch-level fix is; minor would claim a new capability, and the placeholder prop was already declared, typed and defaulted, it simply never reached the DOM. major is forbidden repo-wide (one fixed group). skip-changeset was not used - it is a phantom label here. Gate verdict quoted verbatim: '3 source file(s) of 1 released package(s) changed, and this change declares 1 changeset(s): .changeset/8450-inspector-select-placeholder.md.' re-run green against the new merge-base 61e649a.",
"contradicts_the_brief": [
"THE CALL-SITE COUNT IS 45, NOT 61. Measured non-test JSX elements: 45 at my base c90395b AND 45 at the earlier main tip 7cf6f38, so it is not a figure that decayed - it is not reproducible at any commit I can reach. 51 including test files. The coordinator's message still says 'the 61 call sites'; it should say 45. Coda worth knowing: the repo-wide element count reads exactly 61 AFTER my pins land, because _shared.select.test.tsx goes from 3 to 13 elements. 51 + 10 = 61. If a future reader greps the merged tree and gets 61, that is my test file, not call sites.",
"NO CALLER PASSES A PLACEHOLDER AT ALL - 0 of 45. The card's title, 'no caller's placeholder ever renders', is about a prop nobody uses. What was unreachable is the primitive's own '—' default. The dispatch's sampling instruction ('which call sites pass an explicit placeholder, which rely on the default') therefore has a degenerate answer and could not have chosen a sample; I replaced it with the A/B/C emptiness bucketing above.",
"THE TWO REPAIRS ARE NOT EQUIVALENT ON COST, so the dispatch's 'if they come out equivalent, say so and pick' branch does not apply. A carries a measured controlled-to-uncontrolled flip that fires on data arrival alone.",
"CASE 3 IS NOT THE ABLATION, AND CANNOT BE. Under the ablation it stays GREEN - it must, it is the non-regression half. The ablation is carried by the new positive rows. Case 3 IS load-bearing, but against the always-placeholder caricature, where it goes red. Keeping it was right; calling it the ablation was not.",
"PR #8473 HAD NOT LANDED WHEN I BRANCHED, and it is not 'a sibling file'. It landed as c1ae1af during this run and edited FlowNodeInspector.declaredDefault.test.tsx - the very file this card also had to touch - producing a real merge conflict in its docblock. The dispatch's warning to re-derive line numbers was right for the wrong reason.",
"MY OWN data-placeholder CASE IS THE #8474 SHAPE. It looks like the sharpest case in the file and it survives the always-placeholder caricature untouched. Reported above rather than quietly presented as the strong pin."
],
"tests": "Complete at fbd04bf (pre-merge): 'pnpm exec vitest run packages/app-shell/' Test Files 650 passed (650), Tests 6281 passed | 1 skipped (6282). 'pnpm exec vitest run apps/console/ examples/byo-backend-console/ examples/console-starter/' Test Files 90 passed (90), Tests 1078 passed (1078) - the other three packages TURBO_SCM_BASE=c90395b20 turbo ls --affected reports, run from the repo root with positional paths, never --filter-ed, because a filtered examples/* run relocates process.cwd(). 'pnpm --workspace-concurrency=2 --filter @object-ui/app-shell^... build' exit 0. Complete at the final merged head 7d664d4: 'pnpm exec vitest run packages/app-shell/src/views/metadata-admin/inspectors/' Test Files 69 passed (69), Tests 800 passed | 1 skipped (801) - this is where every InspectorSelectField consumer lives; 'pnpm --filter @object-ui/app-shell run type-check' exit 0 (it is tsc --noEmit && tsc -p tsconfig.test.json, so the new pins are type-checked, and it is what proves data-placeholder on SelectTrigger type-checks); 'pnpm --filter @object-ui/app-shell run lint' (package-scoped eslint ., no narrowing) exit 0, 2925 problems, 0 errors, and 0 hits from main's new no-line-address-in-test-name rule; changeset gate green; check-governed-queue-guard --test on the 4 paths returns NOT GOVERNED; control-byte scan of the 4 files clean at exit 1 with a lit control that fired on a planted 0x01. DECLARED NARROWING, per the PM ruling: the full 'vitest run packages/app-shell/' and the console/examples union were re-launched at 7d664d4 and did NOT complete before this report - the logs carry no summary line and the workers were still alive; they are NOT MEASURED at the merged head and are left to CI. They are measured at fbd04bf, and the only source delta between the two heads is main's own commits plus one docblock paragraph in a test file, with my three product-file blobs byte-identical (_shared.tsx ccdf0f887245e15fd17e16a12752a8582f65f069 at both). ABLATION and CARICATURE: see the two dedicated fields; both proved on disk by two hashes plus both grep directions with matched lines, both restored by state, both under trap EXIT INT TERM with absolute paths. No assertion in the new pins depends on window.innerWidth - nothing opens the Radix content, the only viewport-sensitive part, every reading is the closed trigger's textContent or an attribute, and no responsive class is asserted.",
"mcp_calls": "10 - issue_read get 8450, issue_read get_comments 8450, issue_read get 3912, search_issues (one targeted dedup), issue_write create 8488, issue_read get 8488 (read-back), create_pull_request 8494, pull_request_read get 8494 (full body read-back, intact), add_issue_comment for this report, and this report's own delivery. Card bodies and comments came from MCP; everything else came from git and the local tree.",
"open_questions": [],
"out_of_scope_findings": [
"filed as #8488: InspectorSelectField renders a BLANK trigger for a stored value that matches no option - three call sites (ActionDefaultInspector's ActionTargetField 'not found', ViewColumnInspector's fieldOptions, FlowNodeConfigField's 'deprecated') each hand-roll their own synthesized row, the other 42 do not. Deliberately left out of this PR: an unknown stored value is not an empty one, and showing the placeholder for it would assert 'nothing is stored' about a field that is storing something. Labelled finding + package: app-shell + domain:ui, unassigned, dedup run with a lit control.",
"not filed, reported here: the '61 call sites' figure in this card's title and in the dispatch is unreproducible at any commit (45 non-test at both c90395b and 7cf6f38). No defect, but the title will mislead anyone who greps the merged tree, where 61 is what my own test file makes the total."
]
}

---

Generated by Claude Code

[os-justin]

Correction to this card's own count — "61 call sites" was mine, and it is worse than a stale number

The title said 61. The real figure is 45. I have corrected the title; this comment records why the wrong number needed more than a silent edit.

Measured on origin/main (git grep -o '<InspectorSelectField' -- 'packages/*/src/*'):

tree	<InspectorSelectField elements	production call sites
origin/main	51	45 (6 of the 51 are in _shared.labels.test.tsx / _shared.select.test.tsx)
claude/issue-8450-inspector-select-placeholder (PR #8494)	61	45

61 is not a number from nowhere, and it is not merely stale. It is exactly what the merged tree will read once PR #8494's pins land. The fix PR adds ten more elements — all of them test fixtures — and the repo-wide element count lands on the precise figure this card asserted before any of them existed.

So a future reader who greps the merged tree, gets 61, and concludes the card was right will be counting the fix's own test file as evidence for the claim. A wrong number that becomes self-confirming is worth more than a footnote in a review thread, which is why it goes on the card itself.

The number that matters for the defect is 45: the production call sites at which the placeholder prop was accepted and never reachable.

Credit where it is due — the dev corrected this in the commit title (…unreachable at all 45 call sites) rather than carrying my figure forward.

---

Generated by Claude Code